drivers/xen/events.c

   1 /*
   2  * Xen event channels
   3  *
   4  * Xen models interrupts with abstract event channels.  Because each
   5  * domain gets 1024 event channels, but NR_IRQ is not that large, we
   6  * must dynamically map irqs<->event channels.  The event channels
   7  * interface with the rest of the kernel by defining a xen interrupt
   8  * chip.  When an event is received, it is mapped to an irq and sent
   9  * through the normal interrupt processing path.
  10  *
  11  * There are four kinds of events which can be mapped to an event
  12  * channel:
  13  *
  14  * 1. Inter-domain notifications.  This includes all the virtual
  15  *    device events, since they're driven by front-ends in another domain
  16  *    (typically dom0).
  17  * 2. VIRQs, typically used for timers.  These are per-cpu events.
  18  * 3. IPIs.
  19  * 4. PIRQs - Hardware interrupts.
  20  *
  21  * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
  22  */
  23
  24 #include <linux/linkage.h>
  25 #include <linux/interrupt.h>
  26 #include <linux/irq.h>
  27 #include <linux/module.h>
  28 #include <linux/string.h>
  29 #include <linux/bootmem.h>
  30 #include <linux/slab.h>
  31 #include <linux/irqnr.h>
  32 #include <linux/pci.h>
  33
  34 #include <asm/desc.h>
  35 #include <asm/ptrace.h>
  36 #include <asm/irq.h>
  37 #include <asm/idle.h>
  38 #include <asm/io_apic.h>
  39 #include <asm/sync_bitops.h>
  40 #include <asm/xen/page.h>
  41 #include <asm/xen/pci.h>
  42 #include <asm/xen/hypercall.h>
  43 #include <asm/xen/hypervisor.h>
  44
  45 #include <xen/xen.h>
  46 #include <xen/hvm.h>
  47 #include <xen/xen-ops.h>
  48 #include <xen/events.h>
  49 #include <xen/interface/xen.h>
  50 #include <xen/interface/event_channel.h>
  51 #include <xen/interface/hvm/hvm_op.h>
  52 #include <xen/interface/hvm/params.h>
  53
  54 /*
  55  * This lock protects updates to the following mapping and reference-count
  56  * arrays. The lock does not need to be acquired to read the mapping tables.
  57  */
  58 static DEFINE_MUTEX(irq_mapping_update_lock);
  59
  60 static LIST_HEAD(xen_irq_list_head);
  61
  62 /* IRQ <-> VIRQ mapping. */
  63 static DEFINE_PER_CPU(int [NR_VIRQS], virq_to_irq) = {[0 ... NR_VIRQS-1] = -1};
  64
  65 /* IRQ <-> IPI mapping */
  66 static DEFINE_PER_CPU(int [XEN_NR_IPIS], ipi_to_irq) = {[0 ... XEN_NR_IPIS-1] = -1};
  67
  68 /* Interrupt types. */
  69 enum xen_irq_type {
  70         IRQT_UNBOUND = 0,
  71         IRQT_PIRQ,
  72         IRQT_VIRQ,
  73         IRQT_IPI,
  74         IRQT_EVTCHN
  75 };
  76
  77 /*
  78  * Packed IRQ information:
  79  * type - enum xen_irq_type
  80  * event channel - irq->event channel mapping
  81  * cpu - cpu this event channel is bound to
  82  * index - type-specific information:
  83  *    PIRQ - vector, with MSB being "needs EIO", or physical IRQ of the HVM
  84  *           guest, or GSI (real passthrough IRQ) of the device.
  85  *    VIRQ - virq number
  86  *    IPI - IPI vector
  87  *    EVTCHN -
  88  */
  89 struct irq_info {
  90         struct list_head list;
  91         int refcnt;
  92         enum xen_irq_type type; /* type */
  93         unsigned irq;
  94         unsigned short evtchn;  /* event channel */
  95         unsigned short cpu;     /* cpu bound */
  96
  97         union {
  98                 unsigned short virq;
  99                 enum ipi_vector ipi;
 100                 struct {
 101                         unsigned short pirq;
 102                         unsigned short gsi;
 103                         unsigned char vector;
 104                         unsigned char flags;
 105                         uint16_t domid;
 106                 } pirq;
 107         } u;
 108 };
 109 #define PIRQ_NEEDS_EOI  (1 << 0)
 110 #define PIRQ_SHAREABLE  (1 << 1)
 111
 112 static int *evtchn_to_irq;
 113 static unsigned long *pirq_eoi_map;
 114 static bool (*pirq_needs_eoi)(unsigned irq);
 115
 116 static DEFINE_PER_CPU(unsigned long [NR_EVENT_CHANNELS/BITS_PER_LONG],
 117                       cpu_evtchn_mask);
 118
 119 /* Xen will never allocate port zero for any purpose. */
 120 #define VALID_EVTCHN(chn)       ((chn) != 0)
 121
 122 static struct irq_chip xen_dynamic_chip;
 123 static struct irq_chip xen_percpu_chip;
 124 static struct irq_chip xen_pirq_chip;
 125 static void enable_dynirq(struct irq_data *data);
 126 static void disable_dynirq(struct irq_data *data);
 127
 128 /* Get info for IRQ */
 129 static struct irq_info *info_for_irq(unsigned irq)
 130 {
 131         return irq_get_handler_data(irq);
 132 }
 133
 134 /* Constructors for packed IRQ information. */
 135 static void xen_irq_info_common_init(struct irq_info *info,
 136                                      unsigned irq,
 137                                      enum xen_irq_type type,
 138                                      unsigned short evtchn,
 139                                      unsigned short cpu)
 140 {
 141
 142         BUG_ON(info->type != IRQT_UNBOUND && info->type != type);
 143
 144         info->type = type;
 145         info->irq = irq;
 146         info->evtchn = evtchn;
 147         info->cpu = cpu;
 148
 149         evtchn_to_irq[evtchn] = irq;
 150 }
 151
 152 static void xen_irq_info_evtchn_init(unsigned irq,
 153                                      unsigned short evtchn)
 154 {
 155         struct irq_info *info = info_for_irq(irq);
 156
 157         xen_irq_info_common_init(info, irq, IRQT_EVTCHN, evtchn, 0);
 158 }
 159
 160 static void xen_irq_info_ipi_init(unsigned cpu,
 161                                   unsigned irq,
 162                                   unsigned short evtchn,
 163                                   enum ipi_vector ipi)
 164 {
 165         struct irq_info *info = info_for_irq(irq);
 166
 167         xen_irq_info_common_init(info, irq, IRQT_IPI, evtchn, 0);
 168
 169         info->u.ipi = ipi;
 170
 171         per_cpu(ipi_to_irq, cpu)[ipi] = irq;
 172 }
 173
 174 static void xen_irq_info_virq_init(unsigned cpu,
 175                                    unsigned irq,
 176                                    unsigned short evtchn,
 177                                    unsigned short virq)
 178 {
 179         struct irq_info *info = info_for_irq(irq);
 180
 181         xen_irq_info_common_init(info, irq, IRQT_VIRQ, evtchn, 0);
 182
 183         info->u.virq = virq;
 184
 185         per_cpu(virq_to_irq, cpu)[virq] = irq;
 186 }
 187
 188 static void xen_irq_info_pirq_init(unsigned irq,
 189                                    unsigned short evtchn,
 190                                    unsigned short pirq,
 191                                    unsigned short gsi,
 192                                    unsigned short vector,
 193                                    uint16_t domid,
 194                                    unsigned char flags)
 195 {
 196         struct irq_info *info = info_for_irq(irq);
 197
 198         xen_irq_info_common_init(info, irq, IRQT_PIRQ, evtchn, 0);
 199
 200         info->u.pirq.pirq = pirq;
 201         info->u.pirq.gsi = gsi;
 202         info->u.pirq.vector = vector;
 203         info->u.pirq.domid = domid;
 204         info->u.pirq.flags = flags;
 205 }
 206
 207 /*
 208  * Accessors for packed IRQ information.
 209  */
 210 static unsigned int evtchn_from_irq(unsigned irq)
 211 {
 212         if (unlikely(WARN(irq < 0 || irq >= nr_irqs, "Invalid irq %d!\n", irq)))
 213                 return 0;
 214
 215         return info_for_irq(irq)->evtchn;
 216 }
 217
 218 unsigned irq_from_evtchn(unsigned int evtchn)
 219 {
 220         return evtchn_to_irq[evtchn];
 221 }
 222 EXPORT_SYMBOL_GPL(irq_from_evtchn);
 223
 224 static enum ipi_vector ipi_from_irq(unsigned irq)
 225 {
 226         struct irq_info *info = info_for_irq(irq);
 227
 228         BUG_ON(info == NULL);
 229         BUG_ON(info->type != IRQT_IPI);
 230
 231         return info->u.ipi;
 232 }
 233
 234 static unsigned virq_from_irq(unsigned irq)
 235 {
 236         struct irq_info *info = info_for_irq(irq);
 237
 238         BUG_ON(info == NULL);
 239         BUG_ON(info->type != IRQT_VIRQ);
 240
 241         return info->u.virq;
 242 }
 243
 244 static unsigned pirq_from_irq(unsigned irq)
 245 {
 246         struct irq_info *info = info_for_irq(irq);
 247
 248         BUG_ON(info == NULL);
 249         BUG_ON(info->type != IRQT_PIRQ);
 250
 251         return info->u.pirq.pirq;
 252 }
 253
 254 static enum xen_irq_type type_from_irq(unsigned irq)
 255 {
 256         return info_for_irq(irq)->type;
 257 }
 258
 259 static unsigned cpu_from_irq(unsigned irq)
 260 {
 261         return info_for_irq(irq)->cpu;
 262 }
 263
 264 static unsigned int cpu_from_evtchn(unsigned int evtchn)
 265 {
 266         int irq = evtchn_to_irq[evtchn];
 267         unsigned ret = 0;
 268
 269         if (irq != -1)
 270                 ret = cpu_from_irq(irq);
 271
 272         return ret;
 273 }
 274
 275 static bool pirq_check_eoi_map(unsigned irq)
 276 {
 277         return test_bit(pirq_from_irq(irq), pirq_eoi_map);
 278 }
 279
 280 static bool pirq_needs_eoi_flag(unsigned irq)
 281 {
 282         struct irq_info *info = info_for_irq(irq);
 283         BUG_ON(info->type != IRQT_PIRQ);
 284
 285         return info->u.pirq.flags & PIRQ_NEEDS_EOI;
 286 }
 287
 288 static inline unsigned long active_evtchns(unsigned int cpu,
 289                                            struct shared_info *sh,
 290                                            unsigned int idx)
 291 {
 292         return sh->evtchn_pending[idx] &
 293                 per_cpu(cpu_evtchn_mask, cpu)[idx] &
 294                 ~sh->evtchn_mask[idx];
 295 }
 296
 297 static void bind_evtchn_to_cpu(unsigned int chn, unsigned int cpu)
 298 {
 299         int irq = evtchn_to_irq[chn];
 300
 301         BUG_ON(irq == -1);
 302 #ifdef CONFIG_SMP
 303         cpumask_copy(irq_to_desc(irq)->irq_data.affinity, cpumask_of(cpu));
 304 #endif
 305
 306         clear_bit(chn, per_cpu(cpu_evtchn_mask, cpu_from_irq(irq)));
 307         set_bit(chn, per_cpu(cpu_evtchn_mask, cpu));
 308
 309         info_for_irq(irq)->cpu = cpu;
 310 }
 311
 312 static void init_evtchn_cpu_bindings(void)
 313 {
 314         int i;
 315 #ifdef CONFIG_SMP
 316         struct irq_info *info;
 317
 318         /* By default all event channels notify CPU#0. */
 319         list_for_each_entry(info, &xen_irq_list_head, list) {
 320                 struct irq_desc *desc = irq_to_desc(info->irq);
 321                 cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
 322         }
 323 #endif
 324
 325         for_each_possible_cpu(i)
 326                 memset(per_cpu(cpu_evtchn_mask, i),
 327                        (i == 0) ? ~0 : 0, sizeof(*per_cpu(cpu_evtchn_mask, i)));
 328 }
 329
 330 static inline void clear_evtchn(int port)
 331 {
 332         struct shared_info *s = HYPERVISOR_shared_info;
 333         sync_clear_bit(port, &s->evtchn_pending[0]);
 334 }
 335
 336 static inline void set_evtchn(int port)
 337 {
 338         struct shared_info *s = HYPERVISOR_shared_info;
 339         sync_set_bit(port, &s->evtchn_pending[0]);
 340 }
 341
 342 static inline int test_evtchn(int port)
 343 {
 344         struct shared_info *s = HYPERVISOR_shared_info;
 345         return sync_test_bit(port, &s->evtchn_pending[0]);
 346 }
 347
 348
 349 /**
 350  * notify_remote_via_irq - send event to remote end of event channel via irq
 351  * @irq: irq of event channel to send event to
 352  *
 353  * Unlike notify_remote_via_evtchn(), this is safe to use across
 354  * save/restore. Notifications on a broken connection are silently
 355  * dropped.
 356  */
 357 void notify_remote_via_irq(int irq)
 358 {
 359         int evtchn = evtchn_from_irq(irq);
 360
 361         if (VALID_EVTCHN(evtchn))
 362                 notify_remote_via_evtchn(evtchn);
 363 }
 364 EXPORT_SYMBOL_GPL(notify_remote_via_irq);
 365
 366 static void mask_evtchn(int port)
 367 {
 368         struct shared_info *s = HYPERVISOR_shared_info;
 369         sync_set_bit(port, &s->evtchn_mask[0]);
 370 }
 371
 372 static void unmask_evtchn(int port)
 373 {
 374         struct shared_info *s = HYPERVISOR_shared_info;
 375         unsigned int cpu = get_cpu();
 376         int do_hypercall = 0, evtchn_pending = 0;
 377
 378         BUG_ON(!irqs_disabled());
 379
 380         if (unlikely((cpu != cpu_from_evtchn(port))))
 381                 do_hypercall = 1;
 382         else
 383                 evtchn_pending = sync_test_bit(port, &s->evtchn_pending[0]);
 384
 385         if (unlikely(evtchn_pending && xen_hvm_domain()))
 386                 do_hypercall = 1;
 387
 388         /* Slow path (hypercall) if this is a non-local port or if this is
 389          * an hvm domain and an event is pending (hvm domains don't have
 390          * their own implementation of irq_enable). */
 391         if (do_hypercall) {
 392                 struct evtchn_unmask unmask = { .port = port };
 393                 (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
 394         } else {
 395                 struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
 396
 397                 sync_clear_bit(port, &s->evtchn_mask[0]);
 398
 399                 /*
 400                  * The following is basically the equivalent of
 401                  * 'hw_resend_irq'. Just like a real IO-APIC we 'lose
 402                  * the interrupt edge' if the channel is masked.
 403                  */
 404                 if (evtchn_pending &&
 405                     !sync_test_and_set_bit(port / BITS_PER_LONG,
 406                                            &vcpu_info->evtchn_pending_sel))
 407                         vcpu_info->evtchn_upcall_pending = 1;
 408         }
 409
 410         put_cpu();
 411 }
 412
 413 static void xen_irq_init(unsigned irq)
 414 {
 415         struct irq_info *info;
 416 #ifdef CONFIG_SMP
 417         struct irq_desc *desc = irq_to_desc(irq);
 418
 419         /* By default all event channels notify CPU#0. */
 420         cpumask_copy(desc->irq_data.affinity, cpumask_of(0));
 421 #endif
 422
 423         info = kzalloc(sizeof(*info), GFP_KERNEL);
 424         if (info == NULL)
 425                 panic("Unable to allocate metadata for IRQ%d\n", irq);
 426
 427         info->type = IRQT_UNBOUND;
 428         info->refcnt = -1;
 429
 430         irq_set_handler_data(irq, info);
 431
 432         list_add_tail(&info->list, &xen_irq_list_head);
 433 }
 434
 435 static int __must_check xen_allocate_irq_dynamic(void)
 436 {
 437         int first = 0;
 438         int irq;
 439
 440 #ifdef CONFIG_X86_IO_APIC
 441         /*
 442          * For an HVM guest or domain 0 which see "real" (emulated or
 443          * actual respectively) GSIs we allocate dynamic IRQs
 444          * e.g. those corresponding to event channels or MSIs
 445          * etc. from the range above those "real" GSIs to avoid
 446          * collisions.
 447          */
 448         if (xen_initial_domain() || xen_hvm_domain())
 449                 first = get_nr_irqs_gsi();
 450 #endif
 451
 452         irq = irq_alloc_desc_from(first, -1);
 453
 454         if (irq >= 0)
 455                 xen_irq_init(irq);
 456
 457         return irq;
 458 }
 459
 460 static int __must_check xen_allocate_irq_gsi(unsigned gsi)
 461 {
 462         int irq;
 463
 464         /*
 465          * A PV guest has no concept of a GSI (since it has no ACPI
 466          * nor access to/knowledge of the physical APICs). Therefore
 467          * all IRQs are dynamically allocated from the entire IRQ
 468          * space.
 469          */
 470         if (xen_pv_domain() && !xen_initial_domain())
 471                 return xen_allocate_irq_dynamic();
 472
 473         /* Legacy IRQ descriptors are already allocated by the arch. */
 474         if (gsi < NR_IRQS_LEGACY)
 475                 irq = gsi;
 476         else
 477                 irq = irq_alloc_desc_at(gsi, -1);
 478
 479         xen_irq_init(irq);
 480
 481         return irq;
 482 }
 483
 484 static void xen_free_irq(unsigned irq)
 485 {
 486         struct irq_info *info = irq_get_handler_data(irq);
 487
 488         list_del(&info->list);
 489
 490         irq_set_handler_data(irq, NULL);
 491
 492         WARN_ON(info->refcnt > 0);
 493
 494         kfree(info);
 495
 496         /* Legacy IRQ descriptors are managed by the arch. */
 497         if (irq < NR_IRQS_LEGACY)
 498                 return;
 499
 500         irq_free_desc(irq);
 501 }
 502
 503 static void pirq_query_unmask(int irq)
 504 {
 505         struct physdev_irq_status_query irq_status;
 506         struct irq_info *info = info_for_irq(irq);
 507
 508         BUG_ON(info->type != IRQT_PIRQ);
 509
 510         irq_status.irq = pirq_from_irq(irq);
 511         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
 512                 irq_status.flags = 0;
 513
 514         info->u.pirq.flags &= ~PIRQ_NEEDS_EOI;
 515         if (irq_status.flags & XENIRQSTAT_needs_eoi)
 516                 info->u.pirq.flags |= PIRQ_NEEDS_EOI;
 517 }
 518
 519 static bool probing_irq(int irq)
 520 {
 521         struct irq_desc *desc = irq_to_desc(irq);
 522
 523         return desc && desc->action == NULL;
 524 }
 525
 526 static void eoi_pirq(struct irq_data *data)
 527 {
 528         int evtchn = evtchn_from_irq(data->irq);
 529         struct physdev_eoi eoi = { .irq = pirq_from_irq(data->irq) };
 530         int rc = 0;
 531
 532         irq_move_irq(data);
 533
 534         if (VALID_EVTCHN(evtchn))
 535                 clear_evtchn(evtchn);
 536
 537         if (pirq_needs_eoi(data->irq)) {
 538                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_eoi, &eoi);
 539                 WARN_ON(rc);
 540         }
 541 }
 542
 543 static void mask_ack_pirq(struct irq_data *data)
 544 {
 545         disable_dynirq(data);
 546         eoi_pirq(data);
 547 }
 548
 549 static unsigned int __startup_pirq(unsigned int irq)
 550 {
 551         struct evtchn_bind_pirq bind_pirq;
 552         struct irq_info *info = info_for_irq(irq);
 553         int evtchn = evtchn_from_irq(irq);
 554         int rc;
 555
 556         BUG_ON(info->type != IRQT_PIRQ);
 557
 558         if (VALID_EVTCHN(evtchn))
 559                 goto out;
 560
 561         bind_pirq.pirq = pirq_from_irq(irq);
 562         /* NB. We are happy to share unless we are probing. */
 563         bind_pirq.flags = info->u.pirq.flags & PIRQ_SHAREABLE ?
 564                                         BIND_PIRQ__WILL_SHARE : 0;
 565         rc = HYPERVISOR_event_channel_op(EVTCHNOP_bind_pirq, &bind_pirq);
 566         if (rc != 0) {
 567                 if (!probing_irq(irq))
 568                         printk(KERN_INFO "Failed to obtain physical IRQ %d\n",
 569                                irq);
 570                 return 0;
 571         }
 572         evtchn = bind_pirq.port;
 573
 574         pirq_query_unmask(irq);
 575
 576         evtchn_to_irq[evtchn] = irq;
 577         bind_evtchn_to_cpu(evtchn, 0);
 578         info->evtchn = evtchn;
 579
 580 out:
 581         unmask_evtchn(evtchn);
 582         eoi_pirq(irq_get_irq_data(irq));
 583
 584         return 0;
 585 }
 586
 587 static unsigned int startup_pirq(struct irq_data *data)
 588 {
 589         return __startup_pirq(data->irq);
 590 }
 591
 592 static void shutdown_pirq(struct irq_data *data)
 593 {
 594         struct evtchn_close close;
 595         unsigned int irq = data->irq;
 596         struct irq_info *info = info_for_irq(irq);
 597         int evtchn = evtchn_from_irq(irq);
 598
 599         BUG_ON(info->type != IRQT_PIRQ);
 600
 601         if (!VALID_EVTCHN(evtchn))
 602                 return;
 603
 604         mask_evtchn(evtchn);
 605
 606         close.port = evtchn;
 607         if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 608                 BUG();
 609
 610         bind_evtchn_to_cpu(evtchn, 0);
 611         evtchn_to_irq[evtchn] = -1;
 612         info->evtchn = 0;
 613 }
 614
 615 static void enable_pirq(struct irq_data *data)
 616 {
 617         startup_pirq(data);
 618 }
 619
 620 static void disable_pirq(struct irq_data *data)
 621 {
 622         disable_dynirq(data);
 623 }
 624
 625 int xen_irq_from_gsi(unsigned gsi)
 626 {
 627         struct irq_info *info;
 628
 629         list_for_each_entry(info, &xen_irq_list_head, list) {
 630                 if (info->type != IRQT_PIRQ)
 631                         continue;
 632
 633                 if (info->u.pirq.gsi == gsi)
 634                         return info->irq;
 635         }
 636
 637         return -1;
 638 }
 639 EXPORT_SYMBOL_GPL(xen_irq_from_gsi);
 640
 641 /*
 642  * Do not make any assumptions regarding the relationship between the
 643  * IRQ number returned here and the Xen pirq argument.
 644  *
 645  * Note: We don't assign an event channel until the irq actually started
 646  * up.  Return an existing irq if we've already got one for the gsi.
 647  *
 648  * Shareable implies level triggered, not shareable implies edge
 649  * triggered here.
 650  */
 651 int xen_bind_pirq_gsi_to_irq(unsigned gsi,
 652                              unsigned pirq, int shareable, char *name)
 653 {
 654         int irq = -1;
 655         struct physdev_irq irq_op;
 656
 657         mutex_lock(&irq_mapping_update_lock);
 658
 659         irq = xen_irq_from_gsi(gsi);
 660         if (irq != -1) {
 661                 printk(KERN_INFO "xen_map_pirq_gsi: returning irq %d for gsi %u\n",
 662                        irq, gsi);
 663                 goto out;
 664         }
 665
 666         irq = xen_allocate_irq_gsi(gsi);
 667         if (irq < 0)
 668                 goto out;
 669
 670         irq_op.irq = irq;
 671         irq_op.vector = 0;
 672
 673         /* Only the privileged domain can do this. For non-priv, the pcifront
 674          * driver provides a PCI bus that does the call to do exactly
 675          * this in the priv domain. */
 676         if (xen_initial_domain() &&
 677             HYPERVISOR_physdev_op(PHYSDEVOP_alloc_irq_vector, &irq_op)) {
 678                 xen_free_irq(irq);
 679                 irq = -ENOSPC;
 680                 goto out;
 681         }
 682
 683         xen_irq_info_pirq_init(irq, 0, pirq, gsi, irq_op.vector, DOMID_SELF,
 684                                shareable ? PIRQ_SHAREABLE : 0);
 685
 686         pirq_query_unmask(irq);
 687         /* We try to use the handler with the appropriate semantic for the
 688          * type of interrupt: if the interrupt is an edge triggered
 689          * interrupt we use handle_edge_irq.
 690          *
 691          * On the other hand if the interrupt is level triggered we use
 692          * handle_fasteoi_irq like the native code does for this kind of
 693          * interrupts.
 694          *
 695          * Depending on the Xen version, pirq_needs_eoi might return true
 696          * not only for level triggered interrupts but for edge triggered
 697          * interrupts too. In any case Xen always honors the eoi mechanism,
 698          * not injecting any more pirqs of the same kind if the first one
 699          * hasn't received an eoi yet. Therefore using the fasteoi handler
 700          * is the right choice either way.
 701          */
 702         if (shareable)
 703                 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
 704                                 handle_fasteoi_irq, name);
 705         else
 706                 irq_set_chip_and_handler_name(irq, &xen_pirq_chip,
 707                                 handle_edge_irq, name);
 708
 709 out:
 710         mutex_unlock(&irq_mapping_update_lock);
 711
 712         return irq;
 713 }
 714
 715 #ifdef CONFIG_PCI_MSI
 716 int xen_allocate_pirq_msi(struct pci_dev *dev, struct msi_desc *msidesc)
 717 {
 718         int rc;
 719         struct physdev_get_free_pirq op_get_free_pirq;
 720
 721         op_get_free_pirq.type = MAP_PIRQ_TYPE_MSI;
 722         rc = HYPERVISOR_physdev_op(PHYSDEVOP_get_free_pirq, &op_get_free_pirq);
 723
 724         WARN_ONCE(rc == -ENOSYS,
 725                   "hypervisor does not support the PHYSDEVOP_get_free_pirq interface\n");
 726
 727         return rc ? -1 : op_get_free_pirq.pirq;
 728 }
 729
 730 int xen_bind_pirq_msi_to_irq(struct pci_dev *dev, struct msi_desc *msidesc,
 731                              int pirq, int vector, const char *name,
 732                              domid_t domid)
 733 {
 734         int irq, ret;
 735
 736         mutex_lock(&irq_mapping_update_lock);
 737
 738         irq = xen_allocate_irq_dynamic();
 739         if (irq < 0)
 740                 goto out;
 741
 742         irq_set_chip_and_handler_name(irq, &xen_pirq_chip, handle_edge_irq,
 743                         name);
 744
 745         xen_irq_info_pirq_init(irq, 0, pirq, 0, vector, domid, 0);
 746         ret = irq_set_msi_desc(irq, msidesc);
 747         if (ret < 0)
 748                 goto error_irq;
 749 out:
 750         mutex_unlock(&irq_mapping_update_lock);
 751         return irq;
 752 error_irq:
 753         mutex_unlock(&irq_mapping_update_lock);
 754         xen_free_irq(irq);
 755         return ret;
 756 }
 757 #endif
 758
 759 int xen_destroy_irq(int irq)
 760 {
 761         struct irq_desc *desc;
 762         struct physdev_unmap_pirq unmap_irq;
 763         struct irq_info *info = info_for_irq(irq);
 764         int rc = -ENOENT;
 765
 766         mutex_lock(&irq_mapping_update_lock);
 767
 768         desc = irq_to_desc(irq);
 769         if (!desc)
 770                 goto out;
 771
 772         if (xen_initial_domain()) {
 773                 unmap_irq.pirq = info->u.pirq.pirq;
 774                 unmap_irq.domid = info->u.pirq.domid;
 775                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_unmap_pirq, &unmap_irq);
 776                 /* If another domain quits without making the pci_disable_msix
 777                  * call, the Xen hypervisor takes care of freeing the PIRQs
 778                  * (free_domain_pirqs).
 779                  */
 780                 if ((rc == -ESRCH && info->u.pirq.domid != DOMID_SELF))
 781                         printk(KERN_INFO "domain %d does not have %d anymore\n",
 782                                 info->u.pirq.domid, info->u.pirq.pirq);
 783                 else if (rc) {
 784                         printk(KERN_WARNING "unmap irq failed %d\n", rc);
 785                         goto out;
 786                 }
 787         }
 788
 789         xen_free_irq(irq);
 790
 791 out:
 792         mutex_unlock(&irq_mapping_update_lock);
 793         return rc;
 794 }
 795
 796 int xen_irq_from_pirq(unsigned pirq)
 797 {
 798         int irq;
 799
 800         struct irq_info *info;
 801
 802         mutex_lock(&irq_mapping_update_lock);
 803
 804         list_for_each_entry(info, &xen_irq_list_head, list) {
 805                 if (info->type != IRQT_PIRQ)
 806                         continue;
 807                 irq = info->irq;
 808                 if (info->u.pirq.pirq == pirq)
 809                         goto out;
 810         }
 811         irq = -1;
 812 out:
 813         mutex_unlock(&irq_mapping_update_lock);
 814
 815         return irq;
 816 }
 817
 818
 819 int xen_pirq_from_irq(unsigned irq)
 820 {
 821         return pirq_from_irq(irq);
 822 }
 823 EXPORT_SYMBOL_GPL(xen_pirq_from_irq);
 824 int bind_evtchn_to_irq(unsigned int evtchn)
 825 {
 826         int irq;
 827
 828         mutex_lock(&irq_mapping_update_lock);
 829
 830         irq = evtchn_to_irq[evtchn];
 831
 832         if (irq == -1) {
 833                 irq = xen_allocate_irq_dynamic();
 834                 if (irq == -1)
 835                         goto out;
 836
 837                 irq_set_chip_and_handler_name(irq, &xen_dynamic_chip,
 838                                               handle_edge_irq, "event");
 839
 840                 xen_irq_info_evtchn_init(irq, evtchn);
 841         } else {
 842                 struct irq_info *info = info_for_irq(irq);
 843                 WARN_ON(info == NULL || info->type != IRQT_EVTCHN);
 844         }
 845         irq_clear_status_flags(irq, IRQ_NOREQUEST|IRQ_NOAUTOEN);
 846
 847 out:
 848         mutex_unlock(&irq_mapping_update_lock);
 849
 850         return irq;
 851 }
 852 EXPORT_SYMBOL_GPL(bind_evtchn_to_irq);
 853
 854 static int bind_ipi_to_irq(unsigned int ipi, unsigned int cpu)
 855 {
 856         struct evtchn_bind_ipi bind_ipi;
 857         int evtchn, irq;
 858
 859         mutex_lock(&irq_mapping_update_lock);
 860
 861         irq = per_cpu(ipi_to_irq, cpu)[ipi];
 862
 863         if (irq == -1) {
 864                 irq = xen_allocate_irq_dynamic();
 865                 if (irq < 0)
 866                         goto out;
 867
 868                 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
 869                                               handle_percpu_irq, "ipi");
 870
 871                 bind_ipi.vcpu = cpu;
 872                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
 873                                                 &bind_ipi) != 0)
 874                         BUG();
 875                 evtchn = bind_ipi.port;
 876
 877                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
 878
 879                 bind_evtchn_to_cpu(evtchn, cpu);
 880         } else {
 881                 struct irq_info *info = info_for_irq(irq);
 882                 WARN_ON(info == NULL || info->type != IRQT_IPI);
 883         }
 884
 885  out:
 886         mutex_unlock(&irq_mapping_update_lock);
 887         return irq;
 888 }
 889
 890 static int bind_interdomain_evtchn_to_irq(unsigned int remote_domain,
 891                                           unsigned int remote_port)
 892 {
 893         struct evtchn_bind_interdomain bind_interdomain;
 894         int err;
 895
 896         bind_interdomain.remote_dom  = remote_domain;
 897         bind_interdomain.remote_port = remote_port;
 898
 899         err = HYPERVISOR_event_channel_op(EVTCHNOP_bind_interdomain,
 900                                           &bind_interdomain);
 901
 902         return err ? : bind_evtchn_to_irq(bind_interdomain.local_port);
 903 }
 904
 905 static int find_virq(unsigned int virq, unsigned int cpu)
 906 {
 907         struct evtchn_status status;
 908         int port, rc = -ENOENT;
 909
 910         memset(&status, 0, sizeof(status));
 911         for (port = 0; port <= NR_EVENT_CHANNELS; port++) {
 912                 status.dom = DOMID_SELF;
 913                 status.port = port;
 914                 rc = HYPERVISOR_event_channel_op(EVTCHNOP_status, &status);
 915                 if (rc < 0)
 916                         continue;
 917                 if (status.status != EVTCHNSTAT_virq)
 918                         continue;
 919                 if (status.u.virq == virq && status.vcpu == cpu) {
 920                         rc = port;
 921                         break;
 922                 }
 923         }
 924         return rc;
 925 }
 926
 927 int bind_virq_to_irq(unsigned int virq, unsigned int cpu)
 928 {
 929         struct evtchn_bind_virq bind_virq;
 930         int evtchn, irq, ret;
 931
 932         mutex_lock(&irq_mapping_update_lock);
 933
 934         irq = per_cpu(virq_to_irq, cpu)[virq];
 935
 936         if (irq == -1) {
 937                 irq = xen_allocate_irq_dynamic();
 938                 if (irq == -1)
 939                         goto out;
 940
 941                 irq_set_chip_and_handler_name(irq, &xen_percpu_chip,
 942                                               handle_percpu_irq, "virq");
 943
 944                 bind_virq.virq = virq;
 945                 bind_virq.vcpu = cpu;
 946                 ret = HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
 947                                                 &bind_virq);
 948                 if (ret == 0)
 949                         evtchn = bind_virq.port;
 950                 else {
 951                         if (ret == -EEXIST)
 952                                 ret = find_virq(virq, cpu);
 953                         BUG_ON(ret < 0);
 954                         evtchn = ret;
 955                 }
 956
 957                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
 958
 959                 bind_evtchn_to_cpu(evtchn, cpu);
 960         } else {
 961                 struct irq_info *info = info_for_irq(irq);
 962                 WARN_ON(info == NULL || info->type != IRQT_VIRQ);
 963         }
 964
 965 out:
 966         mutex_unlock(&irq_mapping_update_lock);
 967
 968         return irq;
 969 }
 970
 971 static void unbind_from_irq(unsigned int irq)
 972 {
 973         struct evtchn_close close;
 974         int evtchn = evtchn_from_irq(irq);
 975         struct irq_info *info = irq_get_handler_data(irq);
 976
 977         mutex_lock(&irq_mapping_update_lock);
 978
 979         if (info->refcnt > 0) {
 980                 info->refcnt--;
 981                 if (info->refcnt != 0)
 982                         goto done;
 983         }
 984
 985         if (VALID_EVTCHN(evtchn)) {
 986                 close.port = evtchn;
 987                 if (HYPERVISOR_event_channel_op(EVTCHNOP_close, &close) != 0)
 988                         BUG();
 989
 990                 switch (type_from_irq(irq)) {
 991                 case IRQT_VIRQ:
 992                         per_cpu(virq_to_irq, cpu_from_evtchn(evtchn))
 993                                 [virq_from_irq(irq)] = -1;
 994                         break;
 995                 case IRQT_IPI:
 996                         per_cpu(ipi_to_irq, cpu_from_evtchn(evtchn))
 997                                 [ipi_from_irq(irq)] = -1;
 998                         break;
 999                 default:
1000                         break;
1001                 }
1002
1003                 /* Closed ports are implicitly re-bound to VCPU0. */
1004                 bind_evtchn_to_cpu(evtchn, 0);
1005
1006                 evtchn_to_irq[evtchn] = -1;
1007         }
1008
1009         BUG_ON(info_for_irq(irq)->type == IRQT_UNBOUND);
1010
1011         xen_free_irq(irq);
1012
1013  done:
1014         mutex_unlock(&irq_mapping_update_lock);
1015 }
1016
1017 int bind_evtchn_to_irqhandler(unsigned int evtchn,
1018                               irq_handler_t handler,
1019                               unsigned long irqflags,
1020                               const char *devname, void *dev_id)
1021 {
1022         int irq, retval;
1023
1024         irq = bind_evtchn_to_irq(evtchn);
1025         if (irq < 0)
1026                 return irq;
1027         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1028         if (retval != 0) {
1029                 unbind_from_irq(irq);
1030                 return retval;
1031         }
1032
1033         return irq;
1034 }
1035 EXPORT_SYMBOL_GPL(bind_evtchn_to_irqhandler);
1036
1037 int bind_interdomain_evtchn_to_irqhandler(unsigned int remote_domain,
1038                                           unsigned int remote_port,
1039                                           irq_handler_t handler,
1040                                           unsigned long irqflags,
1041                                           const char *devname,
1042                                           void *dev_id)
1043 {
1044         int irq, retval;
1045
1046         irq = bind_interdomain_evtchn_to_irq(remote_domain, remote_port);
1047         if (irq < 0)
1048                 return irq;
1049
1050         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1051         if (retval != 0) {
1052                 unbind_from_irq(irq);
1053                 return retval;
1054         }
1055
1056         return irq;
1057 }
1058 EXPORT_SYMBOL_GPL(bind_interdomain_evtchn_to_irqhandler);
1059
1060 int bind_virq_to_irqhandler(unsigned int virq, unsigned int cpu,
1061                             irq_handler_t handler,
1062                             unsigned long irqflags, const char *devname, void *dev_id)
1063 {
1064         int irq, retval;
1065
1066         irq = bind_virq_to_irq(virq, cpu);
1067         if (irq < 0)
1068                 return irq;
1069         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1070         if (retval != 0) {
1071                 unbind_from_irq(irq);
1072                 return retval;
1073         }
1074
1075         return irq;
1076 }
1077 EXPORT_SYMBOL_GPL(bind_virq_to_irqhandler);
1078
1079 int bind_ipi_to_irqhandler(enum ipi_vector ipi,
1080                            unsigned int cpu,
1081                            irq_handler_t handler,
1082                            unsigned long irqflags,
1083                            const char *devname,
1084                            void *dev_id)
1085 {
1086         int irq, retval;
1087
1088         irq = bind_ipi_to_irq(ipi, cpu);
1089         if (irq < 0)
1090                 return irq;
1091
1092         irqflags |= IRQF_NO_SUSPEND | IRQF_FORCE_RESUME | IRQF_EARLY_RESUME;
1093         retval = request_irq(irq, handler, irqflags, devname, dev_id);
1094         if (retval != 0) {
1095                 unbind_from_irq(irq);
1096                 return retval;
1097         }
1098
1099         return irq;
1100 }
1101
1102 void unbind_from_irqhandler(unsigned int irq, void *dev_id)
1103 {
1104         free_irq(irq, dev_id);
1105         unbind_from_irq(irq);
1106 }
1107 EXPORT_SYMBOL_GPL(unbind_from_irqhandler);
1108
1109 int evtchn_make_refcounted(unsigned int evtchn)
1110 {
1111         int irq = evtchn_to_irq[evtchn];
1112         struct irq_info *info;
1113
1114         if (irq == -1)
1115                 return -ENOENT;
1116
1117         info = irq_get_handler_data(irq);
1118
1119         if (!info)
1120                 return -ENOENT;
1121
1122         WARN_ON(info->refcnt != -1);
1123
1124         info->refcnt = 1;
1125
1126         return 0;
1127 }
1128 EXPORT_SYMBOL_GPL(evtchn_make_refcounted);
1129
1130 int evtchn_get(unsigned int evtchn)
1131 {
1132         int irq;
1133         struct irq_info *info;
1134         int err = -ENOENT;
1135
1136         if (evtchn >= NR_EVENT_CHANNELS)
1137                 return -EINVAL;
1138
1139         mutex_lock(&irq_mapping_update_lock);
1140
1141         irq = evtchn_to_irq[evtchn];
1142         if (irq == -1)
1143                 goto done;
1144
1145         info = irq_get_handler_data(irq);
1146
1147         if (!info)
1148                 goto done;
1149
1150         err = -EINVAL;
1151         if (info->refcnt <= 0)
1152                 goto done;
1153
1154         info->refcnt++;
1155         err = 0;
1156  done:
1157         mutex_unlock(&irq_mapping_update_lock);
1158
1159         return err;
1160 }
1161 EXPORT_SYMBOL_GPL(evtchn_get);
1162
1163 void evtchn_put(unsigned int evtchn)
1164 {
1165         int irq = evtchn_to_irq[evtchn];
1166         if (WARN_ON(irq == -1))
1167                 return;
1168         unbind_from_irq(irq);
1169 }
1170 EXPORT_SYMBOL_GPL(evtchn_put);
1171
1172 void xen_send_IPI_one(unsigned int cpu, enum ipi_vector vector)
1173 {
1174         int irq = per_cpu(ipi_to_irq, cpu)[vector];
1175         BUG_ON(irq < 0);
1176         notify_remote_via_irq(irq);
1177 }
1178
1179 irqreturn_t xen_debug_interrupt(int irq, void *dev_id)
1180 {
1181         struct shared_info *sh = HYPERVISOR_shared_info;
1182         int cpu = smp_processor_id();
1183         unsigned long *cpu_evtchn = per_cpu(cpu_evtchn_mask, cpu);
1184         int i;
1185         unsigned long flags;
1186         static DEFINE_SPINLOCK(debug_lock);
1187         struct vcpu_info *v;
1188
1189         spin_lock_irqsave(&debug_lock, flags);
1190
1191         printk("\nvcpu %d\n  ", cpu);
1192
1193         for_each_online_cpu(i) {
1194                 int pending;
1195                 v = per_cpu(xen_vcpu, i);
1196                 pending = (get_irq_regs() && i == cpu)
1197                         ? xen_irqs_disabled(get_irq_regs())
1198                         : v->evtchn_upcall_mask;
1199                 printk("%d: masked=%d pending=%d event_sel %0*lx\n  ", i,
1200                        pending, v->evtchn_upcall_pending,
1201                        (int)(sizeof(v->evtchn_pending_sel)*2),
1202                        v->evtchn_pending_sel);
1203         }
1204         v = per_cpu(xen_vcpu, cpu);
1205
1206         printk("\npending:\n   ");
1207         for (i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--)
1208                 printk("%0*lx%s", (int)sizeof(sh->evtchn_pending[0])*2,
1209                        sh->evtchn_pending[i],
1210                        i % 8 == 0 ? "\n   " : " ");
1211         printk("\nglobal mask:\n   ");
1212         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1213                 printk("%0*lx%s",
1214                        (int)(sizeof(sh->evtchn_mask[0])*2),
1215                        sh->evtchn_mask[i],
1216                        i % 8 == 0 ? "\n   " : " ");
1217
1218         printk("\nglobally unmasked:\n   ");
1219         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
1220                 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
1221                        sh->evtchn_pending[i] & ~sh->evtchn_mask[i],
1222                        i % 8 == 0 ? "\n   " : " ");
1223
1224         printk("\nlocal cpu%d mask:\n   ", cpu);
1225         for (i = (NR_EVENT_CHANNELS/BITS_PER_LONG)-1; i >= 0; i--)
1226                 printk("%0*lx%s", (int)(sizeof(cpu_evtchn[0])*2),
1227                        cpu_evtchn[i],
1228                        i % 8 == 0 ? "\n   " : " ");
1229
1230         printk("\nlocally unmasked:\n   ");
1231         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--) {
1232                 unsigned long pending = sh->evtchn_pending[i]
1233                         & ~sh->evtchn_mask[i]
1234                         & cpu_evtchn[i];
1235                 printk("%0*lx%s", (int)(sizeof(sh->evtchn_mask[0])*2),
1236                        pending, i % 8 == 0 ? "\n   " : " ");
1237         }
1238
1239         printk("\npending list:\n");
1240         for (i = 0; i < NR_EVENT_CHANNELS; i++) {
1241                 if (sync_test_bit(i, sh->evtchn_pending)) {
1242                         int word_idx = i / BITS_PER_LONG;
1243                         printk("  %d: event %d -> irq %d%s%s%s\n",
1244                                cpu_from_evtchn(i), i,
1245                                evtchn_to_irq[i],
1246                                sync_test_bit(word_idx, &v->evtchn_pending_sel)
1247                                              ? "" : " l2-clear",
1248                                !sync_test_bit(i, sh->evtchn_mask)
1249                                              ? "" : " globally-masked",
1250                                sync_test_bit(i, cpu_evtchn)
1251                                              ? "" : " locally-masked");
1252                 }
1253         }
1254
1255         spin_unlock_irqrestore(&debug_lock, flags);
1256
1257         return IRQ_HANDLED;
1258 }
1259
1260 static DEFINE_PER_CPU(unsigned, xed_nesting_count);
1261 static DEFINE_PER_CPU(unsigned int, current_word_idx);
1262 static DEFINE_PER_CPU(unsigned int, current_bit_idx);
1263
1264 /*
1265  * Mask out the i least significant bits of w
1266  */
1267 #define MASK_LSBS(w, i) (w & ((~0UL) << i))
1268
1269 /*
1270  * Search the CPUs pending events bitmasks.  For each one found, map
1271  * the event number to an irq, and feed it into do_IRQ() for
1272  * handling.
1273  *
1274  * Xen uses a two-level bitmap to speed searching.  The first level is
1275  * a bitset of words which contain pending event bits.  The second
1276  * level is a bitset of pending events themselves.
1277  */
1278 static void __xen_evtchn_do_upcall(void)
1279 {
1280         int start_word_idx, start_bit_idx;
1281         int word_idx, bit_idx;
1282         int i;
1283         int cpu = get_cpu();
1284         struct shared_info *s = HYPERVISOR_shared_info;
1285         struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
1286         unsigned count;
1287
1288         do {
1289                 unsigned long pending_words;
1290
1291                 vcpu_info->evtchn_upcall_pending = 0;
1292
1293                 if (__this_cpu_inc_return(xed_nesting_count) - 1)
1294                         goto out;
1295
1296 #ifndef CONFIG_X86 /* No need for a barrier -- XCHG is a barrier on x86. */
1297                 /* Clear master flag /before/ clearing selector flag. */
1298                 wmb();
1299 #endif
1300                 pending_words = xchg(&vcpu_info->evtchn_pending_sel, 0);
1301
1302                 start_word_idx = __this_cpu_read(current_word_idx);
1303                 start_bit_idx = __this_cpu_read(current_bit_idx);
1304
1305                 word_idx = start_word_idx;
1306
1307                 for (i = 0; pending_words != 0; i++) {
1308                         unsigned long pending_bits;
1309                         unsigned long words;
1310
1311                         words = MASK_LSBS(pending_words, word_idx);
1312
1313                         /*
1314                          * If we masked out all events, wrap to beginning.
1315                          */
1316                         if (words == 0) {
1317                                 word_idx = 0;
1318                                 bit_idx = 0;
1319                                 continue;
1320                         }
1321                         word_idx = __ffs(words);
1322
1323                         pending_bits = active_evtchns(cpu, s, word_idx);
1324                         bit_idx = 0; /* usually scan entire word from start */
1325                         if (word_idx == start_word_idx) {
1326                                 /* We scan the starting word in two parts */
1327                                 if (i == 0)
1328                                         /* 1st time: start in the middle */
1329                                         bit_idx = start_bit_idx;
1330                                 else
1331                                         /* 2nd time: mask bits done already */
1332                                         bit_idx &= (1UL << start_bit_idx) - 1;
1333                         }
1334
1335                         do {
1336                                 unsigned long bits;
1337                                 int port, irq;
1338                                 struct irq_desc *desc;
1339
1340                                 bits = MASK_LSBS(pending_bits, bit_idx);
1341
1342                                 /* If we masked out all events, move on. */
1343                                 if (bits == 0)
1344                                         break;
1345
1346                                 bit_idx = __ffs(bits);
1347
1348                                 /* Process port. */
1349                                 port = (word_idx * BITS_PER_LONG) + bit_idx;
1350                                 irq = evtchn_to_irq[port];
1351
1352                                 if (irq != -1) {
1353                                         desc = irq_to_desc(irq);
1354                                         if (desc)
1355                                                 generic_handle_irq_desc(irq, desc);
1356                                 }
1357
1358                                 bit_idx = (bit_idx + 1) % BITS_PER_LONG;
1359
1360                                 /* Next caller starts at last processed + 1 */
1361                                 __this_cpu_write(current_word_idx,
1362                                                  bit_idx ? word_idx :
1363                                                  (word_idx+1) % BITS_PER_LONG);
1364                                 __this_cpu_write(current_bit_idx, bit_idx);
1365                         } while (bit_idx != 0);
1366
1367                         /* Scan start_l1i twice; all others once. */
1368                         if ((word_idx != start_word_idx) || (i != 0))
1369                                 pending_words &= ~(1UL << word_idx);
1370
1371                         word_idx = (word_idx + 1) % BITS_PER_LONG;
1372                 }
1373
1374                 BUG_ON(!irqs_disabled());
1375
1376                 count = __this_cpu_read(xed_nesting_count);
1377                 __this_cpu_write(xed_nesting_count, 0);
1378         } while (count != 1 || vcpu_info->evtchn_upcall_pending);
1379
1380 out:
1381
1382         put_cpu();
1383 }
1384
1385 void xen_evtchn_do_upcall(struct pt_regs *regs)
1386 {
1387         struct pt_regs *old_regs = set_irq_regs(regs);
1388
1389         exit_idle();
1390         irq_enter();
1391
1392         __xen_evtchn_do_upcall();
1393
1394         irq_exit();
1395         set_irq_regs(old_regs);
1396 }
1397
1398 void xen_hvm_evtchn_do_upcall(void)
1399 {
1400         __xen_evtchn_do_upcall();
1401 }
1402 EXPORT_SYMBOL_GPL(xen_hvm_evtchn_do_upcall);
1403
1404 /* Rebind a new event channel to an existing irq. */
1405 void rebind_evtchn_irq(int evtchn, int irq)
1406 {
1407         struct irq_info *info = info_for_irq(irq);
1408
1409         /* Make sure the irq is masked, since the new event channel
1410            will also be masked. */
1411         disable_irq(irq);
1412
1413         mutex_lock(&irq_mapping_update_lock);
1414
1415         /* After resume the irq<->evtchn mappings are all cleared out */
1416         BUG_ON(evtchn_to_irq[evtchn] != -1);
1417         /* Expect irq to have been bound before,
1418            so there should be a proper type */
1419         BUG_ON(info->type == IRQT_UNBOUND);
1420
1421         xen_irq_info_evtchn_init(irq, evtchn);
1422
1423         mutex_unlock(&irq_mapping_update_lock);
1424
1425         /* new event channels are always bound to cpu 0 */
1426         irq_set_affinity(irq, cpumask_of(0));
1427
1428         /* Unmask the event channel. */
1429         enable_irq(irq);
1430 }
1431
1432 /* Rebind an evtchn so that it gets delivered to a specific cpu */
1433 static int rebind_irq_to_cpu(unsigned irq, unsigned tcpu)
1434 {
1435         struct evtchn_bind_vcpu bind_vcpu;
1436         int evtchn = evtchn_from_irq(irq);
1437
1438         if (!VALID_EVTCHN(evtchn))
1439                 return -1;
1440
1441         /*
1442          * Events delivered via platform PCI interrupts are always
1443          * routed to vcpu 0 and hence cannot be rebound.
1444          */
1445         if (xen_hvm_domain() && !xen_have_vector_callback)
1446                 return -1;
1447
1448         /* Send future instances of this interrupt to other vcpu. */
1449         bind_vcpu.port = evtchn;
1450         bind_vcpu.vcpu = tcpu;
1451
1452         /*
1453          * If this fails, it usually just indicates that we're dealing with a
1454          * virq or IPI channel, which don't actually need to be rebound. Ignore
1455          * it, but don't do the xenlinux-level rebind in that case.
1456          */
1457         if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_vcpu, &bind_vcpu) >= 0)
1458                 bind_evtchn_to_cpu(evtchn, tcpu);
1459
1460         return 0;
1461 }
1462
1463 static int set_affinity_irq(struct irq_data *data, const struct cpumask *dest,
1464                             bool force)
1465 {
1466         unsigned tcpu = cpumask_first(dest);
1467
1468         return rebind_irq_to_cpu(data->irq, tcpu);
1469 }
1470
1471 int resend_irq_on_evtchn(unsigned int irq)
1472 {
1473         int masked, evtchn = evtchn_from_irq(irq);
1474         struct shared_info *s = HYPERVISOR_shared_info;
1475
1476         if (!VALID_EVTCHN(evtchn))
1477                 return 1;
1478
1479         masked = sync_test_and_set_bit(evtchn, s->evtchn_mask);
1480         sync_set_bit(evtchn, s->evtchn_pending);
1481         if (!masked)
1482                 unmask_evtchn(evtchn);
1483
1484         return 1;
1485 }
1486
1487 static void enable_dynirq(struct irq_data *data)
1488 {
1489         int evtchn = evtchn_from_irq(data->irq);
1490
1491         if (VALID_EVTCHN(evtchn))
1492                 unmask_evtchn(evtchn);
1493 }
1494
1495 static void disable_dynirq(struct irq_data *data)
1496 {
1497         int evtchn = evtchn_from_irq(data->irq);
1498
1499         if (VALID_EVTCHN(evtchn))
1500                 mask_evtchn(evtchn);
1501 }
1502
1503 static void ack_dynirq(struct irq_data *data)
1504 {
1505         int evtchn = evtchn_from_irq(data->irq);
1506
1507         irq_move_irq(data);
1508
1509         if (VALID_EVTCHN(evtchn))
1510                 clear_evtchn(evtchn);
1511 }
1512
1513 static void mask_ack_dynirq(struct irq_data *data)
1514 {
1515         disable_dynirq(data);
1516         ack_dynirq(data);
1517 }
1518
1519 static int retrigger_dynirq(struct irq_data *data)
1520 {
1521         int evtchn = evtchn_from_irq(data->irq);
1522         struct shared_info *sh = HYPERVISOR_shared_info;
1523         int ret = 0;
1524
1525         if (VALID_EVTCHN(evtchn)) {
1526                 int masked;
1527
1528                 masked = sync_test_and_set_bit(evtchn, sh->evtchn_mask);
1529                 sync_set_bit(evtchn, sh->evtchn_pending);
1530                 if (!masked)
1531                         unmask_evtchn(evtchn);
1532                 ret = 1;
1533         }
1534
1535         return ret;
1536 }
1537
1538 static void restore_pirqs(void)
1539 {
1540         int pirq, rc, irq, gsi;
1541         struct physdev_map_pirq map_irq;
1542         struct irq_info *info;
1543
1544         list_for_each_entry(info, &xen_irq_list_head, list) {
1545                 if (info->type != IRQT_PIRQ)
1546                         continue;
1547
1548                 pirq = info->u.pirq.pirq;
1549                 gsi = info->u.pirq.gsi;
1550                 irq = info->irq;
1551
1552                 /* save/restore of PT devices doesn't work, so at this point the
1553                  * only devices present are GSI based emulated devices */
1554                 if (!gsi)
1555                         continue;
1556
1557                 map_irq.domid = DOMID_SELF;
1558                 map_irq.type = MAP_PIRQ_TYPE_GSI;
1559                 map_irq.index = gsi;
1560                 map_irq.pirq = pirq;
1561
1562                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_map_pirq, &map_irq);
1563                 if (rc) {
1564                         printk(KERN_WARNING "xen map irq failed gsi=%d irq=%d pirq=%d rc=%d\n",
1565                                         gsi, irq, pirq, rc);
1566                         xen_free_irq(irq);
1567                         continue;
1568                 }
1569
1570                 printk(KERN_DEBUG "xen: --> irq=%d, pirq=%d\n", irq, map_irq.pirq);
1571
1572                 __startup_pirq(irq);
1573         }
1574 }
1575
1576 static void restore_cpu_virqs(unsigned int cpu)
1577 {
1578         struct evtchn_bind_virq bind_virq;
1579         int virq, irq, evtchn;
1580
1581         for (virq = 0; virq < NR_VIRQS; virq++) {
1582                 if ((irq = per_cpu(virq_to_irq, cpu)[virq]) == -1)
1583                         continue;
1584
1585                 BUG_ON(virq_from_irq(irq) != virq);
1586
1587                 /* Get a new binding from Xen. */
1588                 bind_virq.virq = virq;
1589                 bind_virq.vcpu = cpu;
1590                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_virq,
1591                                                 &bind_virq) != 0)
1592                         BUG();
1593                 evtchn = bind_virq.port;
1594
1595                 /* Record the new mapping. */
1596                 xen_irq_info_virq_init(cpu, irq, evtchn, virq);
1597                 bind_evtchn_to_cpu(evtchn, cpu);
1598         }
1599 }
1600
1601 static void restore_cpu_ipis(unsigned int cpu)
1602 {
1603         struct evtchn_bind_ipi bind_ipi;
1604         int ipi, irq, evtchn;
1605
1606         for (ipi = 0; ipi < XEN_NR_IPIS; ipi++) {
1607                 if ((irq = per_cpu(ipi_to_irq, cpu)[ipi]) == -1)
1608                         continue;
1609
1610                 BUG_ON(ipi_from_irq(irq) != ipi);
1611
1612                 /* Get a new binding from Xen. */
1613                 bind_ipi.vcpu = cpu;
1614                 if (HYPERVISOR_event_channel_op(EVTCHNOP_bind_ipi,
1615                                                 &bind_ipi) != 0)
1616                         BUG();
1617                 evtchn = bind_ipi.port;
1618
1619                 /* Record the new mapping. */
1620                 xen_irq_info_ipi_init(cpu, irq, evtchn, ipi);
1621                 bind_evtchn_to_cpu(evtchn, cpu);
1622         }
1623 }
1624
1625 /* Clear an irq's pending state, in preparation for polling on it */
1626 void xen_clear_irq_pending(int irq)
1627 {
1628         int evtchn = evtchn_from_irq(irq);
1629
1630         if (VALID_EVTCHN(evtchn))
1631                 clear_evtchn(evtchn);
1632 }
1633 EXPORT_SYMBOL(xen_clear_irq_pending);
1634 void xen_set_irq_pending(int irq)
1635 {
1636         int evtchn = evtchn_from_irq(irq);
1637
1638         if (VALID_EVTCHN(evtchn))
1639                 set_evtchn(evtchn);
1640 }
1641
1642 bool xen_test_irq_pending(int irq)
1643 {
1644         int evtchn = evtchn_from_irq(irq);
1645         bool ret = false;
1646
1647         if (VALID_EVTCHN(evtchn))
1648                 ret = test_evtchn(evtchn);
1649
1650         return ret;
1651 }
1652
1653 /* Poll waiting for an irq to become pending with timeout.  In the usual case,
1654  * the irq will be disabled so it won't deliver an interrupt. */
1655 void xen_poll_irq_timeout(int irq, u64 timeout)
1656 {
1657         evtchn_port_t evtchn = evtchn_from_irq(irq);
1658
1659         if (VALID_EVTCHN(evtchn)) {
1660                 struct sched_poll poll;
1661
1662                 poll.nr_ports = 1;
1663                 poll.timeout = timeout;
1664                 set_xen_guest_handle(poll.ports, &evtchn);
1665
1666                 if (HYPERVISOR_sched_op(SCHEDOP_poll, &poll) != 0)
1667                         BUG();
1668         }
1669 }
1670 EXPORT_SYMBOL(xen_poll_irq_timeout);
1671 /* Poll waiting for an irq to become pending.  In the usual case, the
1672  * irq will be disabled so it won't deliver an interrupt. */
1673 void xen_poll_irq(int irq)
1674 {
1675         xen_poll_irq_timeout(irq, 0 /* no timeout */);
1676 }
1677
1678 /* Check whether the IRQ line is shared with other guests. */
1679 int xen_test_irq_shared(int irq)
1680 {
1681         struct irq_info *info = info_for_irq(irq);
1682         struct physdev_irq_status_query irq_status = { .irq = info->u.pirq.pirq };
1683
1684         if (HYPERVISOR_physdev_op(PHYSDEVOP_irq_status_query, &irq_status))
1685                 return 0;
1686         return !(irq_status.flags & XENIRQSTAT_shared);
1687 }
1688 EXPORT_SYMBOL_GPL(xen_test_irq_shared);
1689
1690 void xen_irq_resume(void)
1691 {
1692         unsigned int cpu, evtchn;
1693         struct irq_info *info;
1694
1695         init_evtchn_cpu_bindings();
1696
1697         /* New event-channel space is not 'live' yet. */
1698         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1699                 mask_evtchn(evtchn);
1700
1701         /* No IRQ <-> event-channel mappings. */
1702         list_for_each_entry(info, &xen_irq_list_head, list)
1703                 info->evtchn = 0; /* zap event-channel binding */
1704
1705         for (evtchn = 0; evtchn < NR_EVENT_CHANNELS; evtchn++)
1706                 evtchn_to_irq[evtchn] = -1;
1707
1708         for_each_possible_cpu(cpu) {
1709                 restore_cpu_virqs(cpu);
1710                 restore_cpu_ipis(cpu);
1711         }
1712
1713         restore_pirqs();
1714 }
1715
1716 static struct irq_chip xen_dynamic_chip __read_mostly = {
1717         .name                   = "xen-dyn",
1718
1719         .irq_disable            = disable_dynirq,
1720         .irq_mask               = disable_dynirq,
1721         .irq_unmask             = enable_dynirq,
1722
1723         .irq_ack                = ack_dynirq,
1724         .irq_mask_ack           = mask_ack_dynirq,
1725
1726         .irq_set_affinity       = set_affinity_irq,
1727         .irq_retrigger          = retrigger_dynirq,
1728 };
1729
1730 static struct irq_chip xen_pirq_chip __read_mostly = {
1731         .name                   = "xen-pirq",
1732
1733         .irq_startup            = startup_pirq,
1734         .irq_shutdown           = shutdown_pirq,
1735         .irq_enable             = enable_pirq,
1736         .irq_disable            = disable_pirq,
1737
1738         .irq_mask               = disable_dynirq,
1739         .irq_unmask             = enable_dynirq,
1740
1741         .irq_ack                = eoi_pirq,
1742         .irq_eoi                = eoi_pirq,
1743         .irq_mask_ack           = mask_ack_pirq,
1744
1745         .irq_set_affinity       = set_affinity_irq,
1746
1747         .irq_retrigger          = retrigger_dynirq,
1748 };
1749
1750 static struct irq_chip xen_percpu_chip __read_mostly = {
1751         .name                   = "xen-percpu",
1752
1753         .irq_disable            = disable_dynirq,
1754         .irq_mask               = disable_dynirq,
1755         .irq_unmask             = enable_dynirq,
1756
1757         .irq_ack                = ack_dynirq,
1758 };
1759
1760 int xen_set_callback_via(uint64_t via)
1761 {
1762         struct xen_hvm_param a;
1763         a.domid = DOMID_SELF;
1764         a.index = HVM_PARAM_CALLBACK_IRQ;
1765         a.value = via;
1766         return HYPERVISOR_hvm_op(HVMOP_set_param, &a);
1767 }
1768 EXPORT_SYMBOL_GPL(xen_set_callback_via);
1769
1770 #ifdef CONFIG_XEN_PVHVM
1771 /* Vector callbacks are better than PCI interrupts to receive event
1772  * channel notifications because we can receive vector callbacks on any
1773  * vcpu and we don't need PCI support or APIC interactions. */
1774 void xen_callback_vector(void)
1775 {
1776         int rc;
1777         uint64_t callback_via;
1778         if (xen_have_vector_callback) {
1779                 callback_via = HVM_CALLBACK_VECTOR(XEN_HVM_EVTCHN_CALLBACK);
1780                 rc = xen_set_callback_via(callback_via);
1781                 if (rc) {
1782                         printk(KERN_ERR "Request for Xen HVM callback vector"
1783                                         " failed.\n");
1784                         xen_have_vector_callback = 0;
1785                         return;
1786                 }
1787                 printk(KERN_INFO "Xen HVM callback vector for event delivery is "
1788                                 "enabled\n");
1789                 /* in the restore case the vector has already been allocated */
1790                 if (!test_bit(XEN_HVM_EVTCHN_CALLBACK, used_vectors))
1791                         alloc_intr_gate(XEN_HVM_EVTCHN_CALLBACK, xen_hvm_callback_vector);
1792         }
1793 }
1794 #else
1795 void xen_callback_vector(void) {}
1796 #endif
1797
1798 void __init xen_init_IRQ(void)
1799 {
1800         int i, rc;
1801
1802         evtchn_to_irq = kcalloc(NR_EVENT_CHANNELS, sizeof(*evtchn_to_irq),
1803                                     GFP_KERNEL);
1804         BUG_ON(!evtchn_to_irq);
1805         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1806                 evtchn_to_irq[i] = -1;
1807
1808         init_evtchn_cpu_bindings();
1809
1810         /* No event channels are 'live' right now. */
1811         for (i = 0; i < NR_EVENT_CHANNELS; i++)
1812                 mask_evtchn(i);
1813
1814         pirq_needs_eoi = pirq_needs_eoi_flag;
1815
1816         if (xen_hvm_domain()) {
1817                 xen_callback_vector();
1818                 native_init_IRQ();
1819                 /* pci_xen_hvm_init must be called after native_init_IRQ so that
1820                  * __acpi_register_gsi can point at the right function */
1821                 pci_xen_hvm_init();
1822         } else {
1823                 struct physdev_pirq_eoi_gmfn eoi_gmfn;
1824
1825                 irq_ctx_init(smp_processor_id());
1826                 if (xen_initial_domain())
1827                         pci_xen_initial_domain();
1828
1829                 pirq_eoi_map = (void *)__get_free_page(GFP_KERNEL|__GFP_ZERO);
1830                 eoi_gmfn.gmfn = virt_to_mfn(pirq_eoi_map);
1831                 rc = HYPERVISOR_physdev_op(PHYSDEVOP_pirq_eoi_gmfn_v2, &eoi_gmfn);
1832                 if (rc != 0) {
1833                         free_page((unsigned long) pirq_eoi_map);
1834                         pirq_eoi_map = NULL;
1835                 } else
1836                         pirq_needs_eoi = pirq_check_eoi_map;
1837         }
1838 }